Krea AI releases Krea-2 open-weight image models with 2-second 2K generation

Krea AI released Krea-2-Turbo and Krea-2-Raw as open weights on June 22, a pair of 12-billion-parameter Diffusion Transformer image models, with the Turbo variant generating images at 2K resolution in roughly 2 seconds.

The release matters because few independent labs ship open-weight image models at that size and speed at once. Krea built the Diffusion Transformer architecture for the model family, pairing it with a Qwen 3 VL text encoder and a stack of recent components: Grouped-Query Attention with sigmoid gating, SwiGLU at 4x expansion, and 3D axial rotary position encoding.

The two checkpoints serve different users. Krea-2-Turbo is post-trained and distilled for the roughly two-second generation, while Krea-2-Raw is the undistilled base for users who want to fine-tune from the original weights.

Krea says Krea-2 ranks in the top 10 on the Artificial Analysis text-to-image leaderboard and second among models from independent labs. That ranking is a vendor claim and not independently confirmed, and leaderboard standings shift as new entries arrive.

The weights are open but the terms are not unrestricted. The Krea 2 Community License requires organizations with more than 50 seats to pay for enterprise use and obliges all users to add technical safeguards against generating illegal material, closer to a custom community license than to Apache or MIT.

The release first surfaced on June 22 and gained traction a week later as the open-weight builds climbed community download charts. Whether Krea-2 holds its leaderboard position once outside testers benchmark it will be the signal to watch.